projects/linux·blob master · refs

322 lines · 9.5 KiB

1// SPDX-License-Identifier: GPL-2.0-only
2
3/*
4 * Include rseq.c without _GNU_SOURCE defined, before including any headers, so
5 * that rseq.c is compiled with its configuration, not KVM selftests' config.
6 */
7#undef _GNU_SOURCE
8#include "../rseq/rseq.c"
9#define _GNU_SOURCE
10
11#include <errno.h>
12#include <fcntl.h>
13#include <pthread.h>
14#include <sched.h>
15#include <stdio.h>
16#include <stdlib.h>
17#include <string.h>
18#include <signal.h>
19#include <syscall.h>
20#include <sys/ioctl.h>
21#include <sys/sysinfo.h>
22#include <asm/barrier.h>
23#include <linux/atomic.h>
24#include <linux/rseq.h>
25#include <linux/unistd.h>
26
27#include "kvm_util.h"
28#include "processor.h"
29#include "test_util.h"
30#include "ucall_common.h"
31
32/*
33 * Any bug related to task migration is likely to be timing-dependent; perform
34 * a large number of migrations to reduce the odds of a false negative.
35 */
36#define NR_TASK_MIGRATIONS 100000
37
38static pthread_t migration_thread;
39static cpu_set_t possible_mask;
40static int min_cpu, max_cpu;
41static bool done;
42
43static atomic_t seq_cnt;
44
45static void guest_code(void)
46{
47 for (;;)
48 GUEST_SYNC(0);
49}
50
51static int next_cpu(int cpu)
52{
53 /*
54 * Advance to the next CPU, skipping those that weren't in the original
55 * affinity set. Sadly, there is no CPU_SET_FOR_EACH, and cpu_set_t's
56 * data storage is considered as opaque. Note, if this task is pinned
57 * to a small set of discontigous CPUs, e.g. 2 and 1023, this loop will
58 * burn a lot cycles and the test will take longer than normal to
59 * complete.
60 */
61 do {
62 cpu++;
63 if (cpu > max_cpu) {
64 cpu = min_cpu;
65 TEST_ASSERT(CPU_ISSET(cpu, &possible_mask),
66 "Min CPU = %d must always be usable", cpu);
67 break;
68 }
69 } while (!CPU_ISSET(cpu, &possible_mask));
70
71 return cpu;
72}
73
74static void *migration_worker(void *__rseq_tid)
75{
76 pid_t rseq_tid = (pid_t)(unsigned long)__rseq_tid;
77 cpu_set_t allowed_mask;
78 int i, cpu;
79
80 CPU_ZERO(&allowed_mask);
81
82 for (i = 0, cpu = min_cpu; i < NR_TASK_MIGRATIONS; i++, cpu = next_cpu(cpu)) {
83 CPU_SET(cpu, &allowed_mask);
84
85 /*
86 * Bump the sequence count twice to allow the reader to detect
87 * that a migration may have occurred in between rseq and sched
88 * CPU ID reads. An odd sequence count indicates a migration
89 * is in-progress, while a completely different count indicates
90 * a migration occurred since the count was last read.
91 */
92 atomic_inc(&seq_cnt);
93
94 /*
95 * Ensure the odd count is visible while getcpu() isn't
96 * stable, i.e. while changing affinity is in-progress.
97 */
98 smp_wmb();
99 kvm_sched_setaffinity(rseq_tid, sizeof(allowed_mask), &allowed_mask);
100 smp_wmb();
101 atomic_inc(&seq_cnt);
102
103 CPU_CLR(cpu, &allowed_mask);
104
105 /*
106 * Wait 1-10us before proceeding to the next iteration and more
107 * specifically, before bumping seq_cnt again. A delay is
108 * needed on three fronts:
109 *
110 * 1. To allow sched_setaffinity() to prompt migration before
111 * ioctl(KVM_RUN) enters the guest so that TIF_NOTIFY_RESUME
112 * (or TIF_NEED_RESCHED, which indirectly leads to handling
113 * NOTIFY_RESUME) is handled in KVM context.
114 *
115 * If NOTIFY_RESUME/NEED_RESCHED is set after KVM enters
116 * the guest, the guest will trigger a IO/MMIO exit all the
117 * way to userspace and the TIF flags will be handled by
118 * the generic "exit to userspace" logic, not by KVM. The
119 * exit to userspace is necessary to give the test a chance
120 * to check the rseq CPU ID (see #2).
121 *
122 * Alternatively, guest_code() could include an instruction
123 * to trigger an exit that is handled by KVM, but any such
124 * exit requires architecture specific code.
125 *
126 * 2. To let ioctl(KVM_RUN) make its way back to the test
127 * before the next round of migration. The test's check on
128 * the rseq CPU ID must wait for migration to complete in
129 * order to avoid false positive, thus any kernel rseq bug
130 * will be missed if the next migration starts before the
131 * check completes.
132 *
133 * 3. To ensure the read-side makes efficient forward progress,
134 * e.g. if getcpu() involves a syscall. Stalling the read-side
135 * means the test will spend more time waiting for getcpu()
136 * to stabilize and less time trying to hit the timing-dependent
137 * bug.
138 *
139 * Because any bug in this area is likely to be timing-dependent,
140 * run with a range of delays at 1us intervals from 1us to 10us
141 * as a best effort to avoid tuning the test to the point where
142 * it can hit _only_ the original bug and not detect future
143 * regressions.
144 *
145 * The original bug can reproduce with a delay up to ~500us on
146 * x86-64, but starts to require more iterations to reproduce
147 * as the delay creeps above ~10us, and the average runtime of
148 * each iteration obviously increases as well. Cap the delay
149 * at 10us to keep test runtime reasonable while minimizing
150 * potential coverage loss.
151 *
152 * The lower bound for reproducing the bug is likely below 1us,
153 * e.g. failures occur on x86-64 with nanosleep(0), but at that
154 * point the overhead of the syscall likely dominates the delay.
155 * Use usleep() for simplicity and to avoid unnecessary kernel
156 * dependencies.
157 */
158 usleep((i % 10) + 1);
159 }
160 done = true;
161 return NULL;
162}
163
164static void calc_min_max_cpu(void)
165{
166 int i, cnt, nproc;
167
168 TEST_REQUIRE(CPU_COUNT(&possible_mask) >= 2);
169
170 /*
171 * CPU_SET doesn't provide a FOR_EACH helper, get the min/max CPU that
172 * this task is affined to in order to reduce the time spent querying
173 * unusable CPUs, e.g. if this task is pinned to a small percentage of
174 * total CPUs.
175 */
176 nproc = get_nprocs_conf();
177 min_cpu = -1;
178 max_cpu = -1;
179 cnt = 0;
180
181 for (i = 0; i < nproc; i++) {
182 if (!CPU_ISSET(i, &possible_mask))
183 continue;
184 if (min_cpu == -1)
185 min_cpu = i;
186 max_cpu = i;
187 cnt++;
188 }
189
190 __TEST_REQUIRE(cnt >= 2,
191 "Only one usable CPU, task migration not possible");
192}
193
194static void help(const char *name)
195{
196 puts("");
197 printf("usage: %s [-h] [-u] [-l latency]\n", name);
198 printf(" -u: Don't sanity check the number of successful KVM_RUNs\n");
199 printf(" -l: Set /dev/cpu_dma_latency to suppress deep sleep states\n");
200 puts("");
201 exit(0);
202}
203
204int main(int argc, char *argv[])
205{
206 int r, i, snapshot, opt, fd = -1, latency = -1;
207 bool skip_sanity_check = false;
208 struct kvm_vm *vm;
209 struct kvm_vcpu *vcpu;
210 u32 cpu, rseq_cpu;
211
212 while ((opt = getopt(argc, argv, "hl:u")) != -1) {
213 switch (opt) {
214 case 'u':
215 skip_sanity_check = true;
216 break;
217 case 'l':
218 latency = atoi_paranoid(optarg);
219 break;
220 case 'h':
221 default:
222 help(argv[0]);
223 break;
224 }
225 }
226
227 kvm_sched_getaffinity(0, sizeof(possible_mask), &possible_mask);
228
229 calc_min_max_cpu();
230
231 r = rseq_register_current_thread();
232 TEST_ASSERT(!r, "rseq_register_current_thread failed, errno = %d (%s)",
233 errno, strerror(errno));
234
235 /*
236 * Create and run a dummy VM that immediately exits to userspace via
237 * GUEST_SYNC, while concurrently migrating the process by setting its
238 * CPU affinity.
239 */
240 vm = vm_create_with_one_vcpu(&vcpu, guest_code);
241
242 kvm_pthread_create(&migration_thread, NULL, migration_worker,
243 (void *)(unsigned long)kvm_gettid());
244
245 if (latency >= 0) {
246 /*
247 * Writes to cpu_dma_latency persist only while the file is
248 * open, i.e. it allows userspace to provide guaranteed latency
249 * while running a workload. Keep the file open until the test
250 * completes, otherwise writing cpu_dma_latency is meaningless.
251 */
252 fd = open("/dev/cpu_dma_latency", O_RDWR);
253 TEST_ASSERT(fd >= 0, __KVM_SYSCALL_ERROR("open() /dev/cpu_dma_latency", fd));
254
255 r = write(fd, &latency, 4);
256 TEST_ASSERT(r >= 1, "Error setting /dev/cpu_dma_latency");
257 }
258
259 for (i = 0; !done; i++) {
260 vcpu_run(vcpu);
261 TEST_ASSERT(get_ucall(vcpu, NULL) == UCALL_SYNC,
262 "Guest failed?");
263
264 /*
265 * Verify rseq's CPU matches sched's CPU. Ensure migration
266 * doesn't occur between getcpu() and reading the rseq cpu_id
267 * by rereading both if the sequence count changes, or if the
268 * count is odd (migration in-progress).
269 */
270 do {
271 /*
272 * Drop bit 0 to force a mismatch if the count is odd,
273 * i.e. if a migration is in-progress.
274 */
275 snapshot = atomic_read(&seq_cnt) & ~1;
276
277 /*
278 * Ensure calling getcpu() and reading rseq.cpu_id complete
279 * in a single "no migration" window, i.e. are not reordered
280 * across the seq_cnt reads.
281 */
282 smp_rmb();
283 r = sys_getcpu(&cpu, NULL);
284 TEST_ASSERT(!r, "getcpu failed, errno = %d (%s)",
285 errno, strerror(errno));
286 rseq_cpu = rseq_current_cpu_raw();
287 smp_rmb();
288 } while (snapshot != atomic_read(&seq_cnt));
289
290 TEST_ASSERT(rseq_cpu == cpu,
291 "rseq CPU = %d, sched CPU = %d", rseq_cpu, cpu);
292 }
293
294 if (fd > 0)
295 close(fd);
296
297 /*
298 * Sanity check that the test was able to enter the guest a reasonable
299 * number of times, e.g. didn't get stalled too often/long waiting for
300 * getcpu() to stabilize. A 2:1 migration:KVM_RUN ratio is a fairly
301 * conservative ratio on x86-64, which can do _more_ KVM_RUNs than
302 * migrations given the 1us+ delay in the migration task.
303 *
304 * Another reason why it may have small migration:KVM_RUN ratio is that,
305 * on systems with large low power mode wakeup latency, it may happen
306 * quite often that the scheduler is not able to wake up the target CPU
307 * before the vCPU thread is scheduled to another CPU.
308 */
309 TEST_ASSERT(skip_sanity_check || i > (NR_TASK_MIGRATIONS / 2),
310 "Only performed %d KVM_RUNs, task stalled too much?\n\n"
311 " Try disabling deep sleep states to reduce CPU wakeup latency,\n"
312 " e.g. via cpuidle.off=1 or via -l <latency>, or run with -u to\n"
313 " disable this sanity check.", i);
314
315 kvm_pthread_join(migration_thread, NULL);
316
317 kvm_vm_free(vm);
318
319 rseq_unregister_current_thread();
320
321 return 0;
322}